GitHub (vllm-project/vllm) · Nov 2, 2026 · 10:22 PM

vLLM 0.33 release notes drop CUDA 12.9 builds and remove draft-slot reservation for DFlash/DSpark speculative decoding

Posted Nov 2 with 927 commits, the notes move vLLM to PyTorch 2.14 with CUDA 13.0-only wheels and images, and say DFlash/DSpark drafters no longer reserve token-budget slots, lifting DeepSeek-V4.1-Flash on GB300 from 1,373 to 3,390 output tok/s in the PR's own benchmark. The GitHub tag is still v0.33.0rc3, and Medusa, the MLP speculator and the Ray compiled-graph executor are removed, so check your setup before upgrading.

Read at github.com

More news

View all

KVARK Sparkle 1.0 is an Apache-2.0 Qwen3.8-27B tune that thinks up to 67% less for about the same accuracy

Egzakta AI Lab's first open weight model, posted Nov 2 in BF16, FP8 and GGUF with an MTP head for speculative decoding, cuts median reasoning tokens 34–67% across five benchmarks and answers in 13.7s versus 28.1s for Qwen3.8 27B on one H200, with accuracy within a point except a +5.3 on a 50 problem LiveCodeBench subset. All numbers are the lab's own, and the GGUF runs locally via llama server hf Egzakta/Kvark Sparkle 1.0 27B GGUF:Q4 K M…

Hugging Face (Egzakta) 22 minutes ago

Qodo's Codex plugin pulls org code rules and PR history in, then reviews Codex's changes before you open a PR

Profiled by OpenAI on Nov 2, the Codex marketplace plugin fetches codebase rules and past PR patterns before Codex writes code, then has Qodo review the changes, including cross repo dependencies, and return findings in the same session. It's a customer story rather than a launch post, so there's no pricing or new release detail…

OpenAI 2 hours ago

Maat lets AI agents diagnose Ubuntu from a no-network strict snap, with fixes you approve through polkit

Shown on HN Nov 2 under GPL 3.0, the MCP server reads the journal, units, processes and disks from a strictly confined snap where AppArmor blocks writes and internet sockets, and the model can only propose typed plans like restarting one failed service, which you apply with maatctl and a polkit prompt into a hash chained audit trail. It's a developer preview tested only on Ubuntu 24.04 arm64, and the author's first local model runs ( gpt oss:20b found the root cause 28 of 30 times) cover ten admittedly easy scenarios (Show HN)…

GitHub (bolesawny7/maat) 3 hours ago

OpenAI is merging ChatGPT's Chat and Work tabs into one, with Codex staying separate

A Nov 2 email to users says one combined view will let you go from talking through an idea to having ChatGPT build a site, make a deck or act on the web with your files and plugins, starting with a small test group before a wider rollout. OpenAI's Thibault Sottiaux said on X it's "not rolling out just yet," and no date has been given (The Verge)…

TechCrunch 4 hours ago

AgentTM records Codex runs with workspace checkpoints you can diff and fork into separate worktrees

Shown on HN Nov 2, the local Go CLI wraps agenttm run codex "task" , logs events to SQLite with before and after workspace snapshots, and serves a UI at 127.0.0.1:7331 to browse historical files, compare diffs and fork any checkpoint into .agenttm/worktrees/ without touching your tree, index or HEAD. v0.1 only captures non interactive Codex runs it launches, file snapshots aren't redacted, and only Windows end to end use is reported as tested…

GitHub (Jyotishmoy12/AgentTM) 5 hours ago

Spotted something we missed? Start a thread.